inclusionAI Open-Sources "Ling-2.6-flash," an Agent-Optimized Model

On May 3, inclusionAI released an open-source version of Ling-2.6-flash, an instruct model optimized for agent use cases, on Hugging Face. It adopts a MoE architecture with 7.4B active parameters out of 104B total, designed to curb the heavy token consumption that becomes a problem in agent workflows.

While the industry mainstream pushes performance ceilings through "long-reasoning," Ling-2.6-flash deliberately focuses its optimization on three things: inference efficiency, token efficiency, and agent performance. It can be seen as a model that answers the question of whether excessive reasoning tokens are truly necessary for everyday, high-frequency agent use cases.

As cost becomes the biggest barrier to running agents in production, this approach—tackling token efficiency head-on—offers high practical value. We look forward to detailed benchmarks in future evaluations.

Chinese Court Rules That "Workers Cannot Be Replaced by AI"

Futurism reports that a court in China has ruled that workers cannot be replaced by AI. Specifically, the court found that it is not legally permissible for a company to fire employees and replace them with AI.

As AI adoption advances, its interplay with employment remains under debate in many countries. While China promotes the aggressive use of AI technology, this ruling could become an important precedent from a worker-protection standpoint. Where to draw the line between "technological innovation" and "workers' rights" is a shared challenge for many countries, including Japan.

For now, it appears to be a ruling on an individual case, but if similar decisions accumulate, it may become a model case for legal frameworks in countries pushing AI adoption forward.

Hummingbird+: LLM Inference on Low-Cost FPGAs, Aiming for $150 at Mass Production

What is making waves in the LocalLLaMA community is Hummingbird+, an FPGA-based approach to LLM inference. It runs a Q4-quantized Qwen3-30B-A3B model at 18 tok/s using 24GB of memory, with mass-production costs estimated at $150.

As an inference method that does not depend on NVIDIA GPUs, FPGAs are drawing attention as a "third path" following AMD and Intel accelerators. Particularly for edge devices and small-scale on-premises deployments, GPU procurement costs and power consumption often become obstacles, and FPGAs could be a compelling option.

It is still at the academic paper stage, and actual productization is expected to take time, but the diversification of LLM inference hardware is a welcome direction.

Lilly × Profluent: $2.25B Mega-Deal in AI Drug Discovery

BioSpace reports that Eli Lilly and AI biotech Profluent have struck a $2.25 billion partnership in a field described as the "holy grail" of gene therapy.

Investment in AI drug discovery has expanded rapidly over the past few years, with AI credited with major contributions, especially in protein design and gene therapy. Profluent holds technology for designing proteins with AI, and this large-scale partnership signals the commercial maturity of the field.

AI drug discovery startups are also on the rise in Japan—a global trend worth watching.

Most Modern Phishing Campaigns Leverage AI — The Register

According to The Register, the majority of modern phishing campaigns make use of AI. The spread of generative AI has made it easy to mass-produce persuasive phishing emails in natural language, and an increasing number of them slip past traditional spam filters.

Security measures now require not only "technical defenses" but also the literacy to spot AI-generated text. Corporate security education faces new demands as well.

Quantized GGUF Version of LLM-jp-4 32B Thinking Released

A GGUF version of the LLM-jp project's llm-jp-4-32b-a3b-thinking (a 32B total / 3B active MoE with chain-of-thought output support), quantized with imatrix calibration using the model's own pre-training corpus, llm-jp-corpus-v4, has been released on Hugging Face.

Its key feature is minimizing quantization accuracy degradation by using the same training corpus as the model rather than a generic one. For users who want to try local Japanese LLM inference, this adds a practical new option.

---

**References**

[Ling-2.6-flash – Hugging Face](https://huggingface.co/inclusionAI/Ling-2.6-flash)

[Chinese Court Rules That a Worker Cannot Be Replaced by AI – Futurism](https://futurism.com/artificial-intelligence/china-legal-ai-automation)

[Hummingbird+: low-cost FPGAs for LLM inference – Reddit/LocalLLaMA](https://www.reddit.com/r/LocalLLaMA/comments/1t2kpzn/paper_on_hummingbird_lowcost_fpgas_for_llm/)

[Lilly, AI biotech Profluent ink $2.25B pact – BioSpace](https://www.biospace.com/deals/lilly-ai-biotech-profluent-ink-2-25b-pact-in-search-of-genetic-medicine-holy-grail)

[Bot her emails: most modern phishing campaigns are AI-enabled – The Register](https://www.theregister.com/2026/04/30/modern_phishing_campaigns_ai/)

[LLM-jp-4 32B Thinking GGUF – Zenn](https://zenn.dev/suzumura_lab/articles/0a1bdb04ec87ca)